Sherly Boddu and Sam Hsu
4/21/2020
All analysis conducted using R 3.6.3.
Visualizations: geographic distribution, density plots, Kaplan-Meier plots
Cox proportional hazard regression:
The hazard function is the instantaneous risk of death at time \(t\), conditional on survival up to that time: \[h(t) = \lim_{\Delta t\rightarrow0} \frac{Pr[(t \leq T < t + \Delta t | T \geq t)]}{\Delta t}\]
Cox proportional-hazards model for relationship of survival distribution to covariates:
\[h_{i}(t) = h_{0}(t) \exp(\alpha + \beta_{sex}x_{i,sex} + \beta_{age}x_{i, age} + \beta_{province}x_{i, province})\]
We assume proportional hazards. This is acceptable, using the cox.zph function to test.
We assume no influential cases. This does not hold at the province level, but in certain provinces only 1 or 2 deaths were recorded.
| Table 1: Characteristics of people diagnosed with COVID-19 in South Korea between 01/20/2020 and 04/20/2020 | Full sample (N = 2,772) | Deaths only (N = 65) |
|---|---|---|
| Sex: | ||
| Male | 1,175 (43.3%) | 44 (65.7%) |
| Female | 1,536 (56.7%) | 23 (34.3%) |
| Age: | ||
| Mean (SD) | 45.53 (20.68) | 74.61 (13.08) |
| Province: | ||
| Busan | 129 (4.7%) | 3 (4.5%) |
| Daegu | 63 (2.3%) | 20 (29.9%) |
| Gangwon-do | 37 (1.3%) | 1 (1.5%) |
| Gyeonggi-do | 656 (23.6%) | 1 (1.5%) |
| Gyeongsangbuk-do | 1,223 (44.1%) | 39 (58.2%) |
| Seoul | 624 (22.5%) | 2 (3.0%) |
| Ulsan | 42 (1.5%) | 1 (1.5%) |
| Days to Resolution (Death or Release): | ||
| Mean (SD) | 21.16 (9.75) | 8.15 (10.10) |
| Disease State: | ||
| Deceased | 67 (2.4%) | 67 (100.0%) |
| Isolated | 1,438 (51.8%) | 0 (0.0%) |
| Released | 1,269 (45.7%) | 0 (0.0%) |
| Characteristic | HR1 | 95% CI1 | p-value |
|---|---|---|---|
| Sex | |||
| female | — | — | |
| male | 3.33 | 1.99, 5.57 | <0.001 |
| Age | 1.08 | 1.06, 1.10 | <0.001 |
| Province | |||
| Busan | — | — | |
| Daegu | 11.3 | 3.30, 38.4 | <0.001 |
| Gangwon-do | 1.52 | 0.16, 14.7 | 0.7 |
| Gyeonggi-do | 0.08 | 0.01, 0.73 | 0.026 |
| Gyeongsangbuk-do | 1.10 | 0.34, 3.62 | 0.9 |
| Seoul | 0.21 | 0.03, 1.26 | 0.087 |
| Ulsan | 2.13 | 0.22, 20.8 | 0.5 |
|
1
HR = Hazard Ratio, CI = Confidence Interval
|
|||
In our study locations, 2.4% of those diagnosed with COVID-19 have died.
There are statistically significant associations between sex and age and time to event.
There are statistically significant associations by province as well, although we caution that due to very low death counts it is difficult to draw any conclusions.
Daegu’s Shincheonji Church made international news for its role in worsening the spread of COVID-19 in South Korea by reportedly hiding names of members.
Early results from China and Italy also suggest that age and sex are risk factors for mortality. (Wu and McGoogan, 2020; Jordan, Adab, and Cheng, 2020; Livingston and Bucher, 2020)
Strengths
South Korean data collection systems are very strong; contact tracing mechanisms suggest reliable data.
To our knowledge, no survival analysis published on COVID-19 cases in South Korea.
Limitations
Low death counts are good! But they make drawing inferences about survival from covariates challenging.
Important unobserved variables: smoking status, prior chronic disease, secondary infection.
Province is not a fine-enough level of geographic analysis.